推荐引擎没有让信息变窄（腾讯科技乌镇专访）

Recommendation Engines Haven't Narrowed Information (Tencent Tech Interview in Wuzhen)

> Outlet: Tencent Tech (text transcript of a video interview during the World Internet Conference in Wuzhen) · 2014-11-19 · [Original link](http://tech.qq.com/a/20141119/041386.htm) ([archived](https://web.archive.org/web/20141129035943/http://tech.qq.com/a/20141119/041386.htm))
> Only his own answers are kept. Prompts marked (Q) are the editor's one-line summaries, not the interviewer's wording.

**(Q)** Toutiao's latest operating figures, including active users?

In the two-plus years since Toutiao launched, it now has 180 million cumulative installed users, nearly 20 million daily active users and 18 million monthly active users. [Translator's note: the figures are as in the source; the reporters' lead gives "50 million monthly active users", and the collector left both unchanged.]

**(Q)** How is revenue?

It's not convenient to release it in great detail, but it has reached our expectations, and progress is good.

**(Q)** Any new ideas on the business model?

Our business model has always been fairly simple and clear. Toutiao is different from other reading products: since the first day it launched, it has never for a second let go of this single-page ad, and it has never done other things like charging users for reading or game apps. We mainly recommend commercially valuable information and in-feed ads to users. We're trying out how different types of industries perform in the feed, so from what we see now this also meets our expectations, and revenue has kept growing steadily; it's quite good.

**(Q)** Will personalized recommendation narrow people's information channels?

Actually, the essence of this question is that people have very high expectations of recommendation engines. When we use a traditional news app, we don't feel it narrows our information channels, but with recommendation engines this worry arises.

Actually I think news apps made my information quite narrow: basically a lot of what I used to read was news about Libya and Pakistan. And it's not only a question of narrowing: traditional news apps, first, don't match users' interests, and second, aren't broad either. Why? What they report is probably always one editor-in-chief's ideas, what he likes, so the slant and the areas of focus are very obvious. Coming back to recommendation engines, people's expectations of them are much higher, and of course I think that expectation is right and good.

Looking at the past, recommendation engines did prioritize content users were interested in. Actually they balance three things. One is inferring from history, picking the topics users are most interested in and care most about. Second, public topics, content the public cares about: you may not be interested in it, but when others mention it, you can't say you don't know. The third kind is in fact the personalized content we're focusing on developing.

Between people there is generalization, and social connection is one kind of generalization. [Translator's note: the source writes 放化 throughout; the collector notes it is probably a mishearing of 泛化, "generalization", which is how his forum speech on the same day is transcribed.] What we call generalization means what is passed between people who have something in common. To explain further: we have friends in common, a geographic location in common, an occupational background in common, an employer in common; these are also kinds of commonality.

Social connection is an important path of generalization, and it's also what we're focusing on now. We've already had very clear results: after seeing diversity improve, user time spent and retention have both made good progress. People who don't understand the technology will say personalization just recommends content based on a user's behavior, but that's actually wrong. First, recommending content based only on the user's own behavior will actually tend to be narrow; second, there isn't enough data either. In fact it's "all for me, me for all": everyone recommends for everyone. So with a good recommendation system, the more people use it, the better it works, and the longer you use it, the better it works; it will help you broaden your information.

**(Q)** Will recommendation algorithms get smarter and smarter?

Yes, because the system knows that someone who likes home furnishings may later like travel, and someone who reads about preparing for pregnancy may go on to read about parenting; it collects statistics on how different groups' interests shift. The ways and paths by which the interests of people with information similar to yours, or with social ties to you, or with the same occupational background, shift are also a path for gauging yours, and they help information exploration. So internally we have a research direction devoted to interest exploration.

**(Q)** Where does user data come from?

The data is mainly users' own behavioral data.

**(Q)** Only from users' behavior on Toutiao?

Yes, on this platform alone the data is already very rich. In addition, we have a lot of cooperation with various social networking sites. If a user logs in with a Weibo account, the system will also explore information from it. [Translator's note: the source sentence breaks off at 信息的探; translated by its apparent sense.] Beyond that, we'll also do some cooperation with other products. But for now data accumulation is still mainly internal.

**(Q)** What do you think of location-based recommendation?

Yes, scenario. Location is only one kind of scenario: location is a scenario, and weather and time are scenarios too. In different scenarios, the content you want to read is different. If you open Toutiao now and refresh it here, there should be more content about the World Internet Conference. Scenario is an important recommendation feature.

This scenario even includes whether you're online over 3G, 4G or 2G.

**(Q)** Does it go down to the device?

It does. For example, if you're in a big-screen, high-bandwidth scenario, you'd rather see content with large pictures and many pictures.

**(Q)** Have you achieved fine-grained recommendation already?

Yes. If you're on a 2G network and we recommend content with large pictures and many pictures, you may get very angry, because it uses so much bandwidth. Combining people, information, scenario and other dimensions and levels can help users discover information in future, and there's still a lot of room for improvement.

**(Q)** How do you achieve win-win with traditional media?

So far we're probably the mobile product with the most media partners; we've reached cooperation with more than 5,000 media organizations. We don't produce content directly, but I think we should communicate and cooperate better with the media that produce content, including in distributing traffic and supporting monetization. This year we've sped up our work on supporting monetization and also done more on traffic distribution, so that media feel a sense of ownership over their users on the platform. We let users become fans of media outlets, added operating functions and added permissions for running ads; all of this has greatly improved our relationship with the media. Now, on mobile, we're the product with the broadest cooperation.

**(Q)** How do you help traditional media monetize?

Through an online ad network. We also cooperate with some media who are making more attempts, for example through ads and services. Now we're doing more of this kind of cooperation.

**(Q)** Any concrete examples?

These attempts may not be very typical. For example, some financial media can, besides reporting financial knowledge, also promote products. For example, there's currently a fashion media account on the platform that, while writing fashion content, also recommends fashion brands' products, and it works quite well. I think these all count as good attempts.

**(Q)** Will you do machine writing?

Yes, robot writing. We think it's possible in specific fields: finance, stocks, weather; these are possible. But highly creative content creation can't be done. I think content with the strongest views and the most brilliant ideas still has to be created by people. Machine-produced content is more about turning raw information into a readable piece; it's more a conversion of format. It can't distill more complex views or do more complex real-time organization. Because machines can't do that yet, I think that part is more for specific fields: formatted information, turned into readable information.

**(Q)** Will you try it in specific fields?

We're making this attempt now. Actually we're making some attempts in weather; we'll make some attempts, but we don't think this is the top priority right now. The top priorities right now are still two things. First, recommend information better, meeting users' need to get information. Second, make the creation and monetization platform better, serving our partner media. We now call it "Toutiao Hao": to support creation and monetization on the Toutiao Hao platform better, so that people who create content create more and better content. I think it's these two things.

**(Q)** Any plans to go public?

Right, because we're an independent company, there's no such thing as a "standalone" listing; a listing is just a listing, there's no "standalone" listing. And we think listing is the path from a private company to a public one, which certainly has to be considered at some point in the future. But we're not in a particular hurry about it. We think what matters most is what helps us do things well, make the product well, and serve our broad user base. So on this, we think if the timing is right, and also this will take up a lot of energy... If the timing is right and our energy allows, we'll consider it, but we're not in a particular hurry, so in the short term it shouldn't happen.

**(Q)** Does the short term mean three to five years?

No, the short term here should be shorter. In the short term it shouldn't happen; over a longer time, I think it depends on the company's situation.

**(Q)** How many employees, and how many work on recommendation?

On optimizing the recommendation engine: about 50% of our employees work directly on product and technology, directly serving product and technology, and the rest are in other functions. Headcount has kept growing fast; we have a few hundred people now, three hundred-some, still growing fast. And we'll keep increasing our investment in basic product technology.

**(Q)** Direction or vision for 2015?

Toutiao's vision has actually never changed. In plain words, we want to help more and more users get rid of the regret "if only I'd known". In its most complete form: we want to help users discover valuable information more effectively. The work we're doing now, whether content partnerships, encouraging people on the platform to create content, or supporting recommendation on more features, is all to let users discover more information, let information reach users earlier, and get rid of the regret "if only I'd known". I think from the founding until now, this has always been simple and clear.

**(Q)** So in future you'll still center on making a good product?

Yes. Everything else, whether listing or other commercial monetization, I think is secondary.

**(Q)** Have you reached profitability?

We're still investing on a large scale; because investment keeps increasing, there's no profit at scale yet.

**(Q)** Any plans for a new round of financing?

Not for now. We still have plenty of money.

Copyright: © LookMinds. Our own edited text and translation, not a substitute for the original; rights in the original belong to the author and the platform. Please see the original link.